Deep Transfer Learning for Person Re-identification (1611.05244v2)

Abstract: Person re-identification (Re-ID) poses a unique challenge to deep learning: how to learn a deep model with millions of parameters on a small training set of few or no labels. In this paper, a number of deep transfer learning models are proposed to address the data sparsity problem. First, a deep network architecture is designed which differs from existing deep Re-ID models in that (a) it is more suitable for transferring representations learned from large image classification datasets, and (b) classification loss and verification loss are combined, each of which adopts a different dropout strategy. Second, a two-stepped fine-tuning strategy is developed to transfer knowledge from auxiliary datasets. Third, given an unlabelled Re-ID dataset, a novel unsupervised deep transfer learning model is developed based on co-training. The proposed models outperform the state-of-the-art deep Re-ID models by large margins: we achieve Rank-1 accuracy of 85.4\%, 83.7\% and 56.3\% on CUHK03, Market1501, and VIPeR respectively, whilst on VIPeR, our unsupervised model (45.1\%) beats most supervised models.

• The paper introduces a dual-loss deep network architecture that integrates classification and verification losses to transfer knowledge effectively.
• It employs a novel two-step fine-tuning strategy to adapt models from large-scale datasets to small Re-ID sets, reducing overfitting and boosting performance.
• Empirical results demonstrate significant gains, with Rank-1 accuracies of 85.4%, 83.7%, and 56.3% on CUHK03, Market1501, and VIPeR benchmarks respectively.

An Analysis of "Deep Transfer Learning for Person Re-identification"

In the paper "Deep Transfer Learning for Person Re-identification," the authors address the complex problem of re-identifying individuals across non-overlapping camera views. Person re-identification (Re-ID) is critical in surveillance systems, and it has posed a unique challenge to the deep learning community due to the sparse availability of training labels. This paper presents a suite of deep transfer learning models designed to overcome data scarcity, by harnessing previously learned representations from large image datasets like ImageNet.

Firstly, the paper proposes an innovative deep network architecture that integrates classification and verification losses with differentiated dropout strategies. This configuration not only assists in transferring knowledge from large image classification tasks but also mitigates overfitting. The architecture employs a base structure derived from GoogLeNet, which combines the ability to classify identities with the subtleties of instance verification. The authors substantiate that this dual-loss setup bridges domain discrepancies inherent between object categorization tasks and instance recognition, while simultaneously addressing domain gaps between auxiliary image datasets and Re-ID tasks.

The paper also introduces a novel two-stepped fine-tuning strategy, which significantly boosts model adaptability from auxiliary datasets to explicit Re-ID datasets. This approach is particularly advantageous when the target Re-ID datasets are small, requiring knowledge transfer from larger datasets like Market1501 or CUHK03. Through this staged transfer learning strategy, the authors illustrate the importance of not only employing ImageNet as a pre-training source but also refining the models with intermediate large Re-ID datasets.

Numerically, the models proposed in the paper achieve substantial performance improvements over state-of-the-art deep Re-ID models. Notably, Rank-1 accuracy on benchmarks such as CUHK03, Market1501, and VIPeR datasets reach 85.4%, 83.7%, and 56.3% respectively. These results validate the models' superiority, indicating a robust capability to generalize across varied Re-ID datasets despite the limited number of labels.

In terms of implications, this research underscores the paradigm shift in training deep Re-ID models. Traditional deep networks, when trained on Re-ID datasets alone, do not fully leverage the available transferable features from more substantial datasets. By effectively transferring knowledge from large-scale image classification datasets, the proposed models set a benchmark for resolving data sparsity in Re-ID applications. This naturally extends into real-world applications where labeling efforts are minimized, yet sophisticated Re-ID systems can still operate effectively.

The unsupervised deep model presented is of particular interest for future developments in AI. By co-training a regular deep model with an unsupervised dictionary learning model, this research opens the door to devising Re-ID systems capable of functioning with unlabeled target datasets. This significantly reduces the dependency on annotations, thus facilitating scalable and practical deployment in diverse surveillance scenarios.

Looking forward, the promising results advocate for further exploration into combining transfer learning with unsupervised learning paradigms. Future work could investigate optimizing deep network architectures specifically for unsupervised domain adaptation, tackling discrepancies not just in feature space but also in data distribution. Additionally, the exploration of more complex and larger auxiliary datasets beyond ImageNet might reveal deeper insights into the robustness and adaptability of deep Re-ID models.

This paper represents a significant step towards solving the person re-identification problem in constrained data environments. It redefines how auxiliary data can be harnessed to accelerate learning, suggesting novel pathways for both academic inquiry and practical application in the ever-growing domain of computer vision.